DeepSeek Harness plugin · Chromium extension
Augmentor lets your agent do the web work for you: it opens pages, reads them, clicks, and types in your real Chromium browser — so you stop doing repetitive tasks by hand. It runs on the DSH you already have, and every action happens behind a frost veil you can watch.
the real side panel · live in your browser
MODELS
The panel’s model picker is the DSH catalog: the same $DSH_HOME/settings.yaml your app reads. Point Augmentor at a 27B on your own GPU or DeepSeek in the cloud, and switch per conversation. No per-route config, no restart.
Use the models already running on your machine through llama.cpp. Private, offline, and as free as your own hardware. Augmentor reads your local routes straight from DSH settings. Nothing extra to wire up.
Flip to the built-in DeepSeek route for frontier models. No local GPU required. The key resolves per request from your DSH credential store, the same one the app’s Models page writes.
IN ACTION
Here’s a turn. On the left, the conversation; on the right, the frost veil on the very tab the agent is driving. You see every step in plain words.
HOW IT WORKS
The extension talks to your running DSH through the native-messaging host. The whole path stays on loopback and reuses the transport you already trust.
01 · EXTENSION
The chat UI, session + model picker, and the per-tab browser executor. It owns the frost veil and status pill.
02 · PROTOCOL PIPE
A small loopback Node process re-frames the port traffic as /api calls + downlink sockets. This is the only process Augmentor owns.
03 · YOUR DSH
The @deepseek-ai/dsh-augmentor plugin registers the browser_* tools and bridges each to the extension. Everything else (model catalog, sessions, Stop) is stock DSH.
FEATURES
Everything below is stock DeepSeek Harness surfaced through the panel. No config you have to re-learn.
The picker is your DSH catalog: local llama.cpp routes and the built-in DeepSeek route, exactly as configured. Switch per conversation, no restart.
browser_tabs_list, browser_navigate, browser_snapshot, browser_click, browser_type, registered in-process on the agent. No bash round-trip, no CLI, no relay.
Conversations are real DSH sessions (session.create), shown in the app sidebar, live in both places. Ephemeral by default, one tap to Save to Augmentor Chat.
Stock session.cancel aborts the live turn while preserving the pending inbox. No local DSH patch, no interrupt route to maintain.
Replies render as they’re generated. The panel coalesces deltas per animation frame with a blinking caret, mirroring the DSH web GUI.
Jump from the panel straight into the running app with the right chat. Saved chats land in Augmentor Chat; the rest stay in Ungrouped, where housekeeping keeps them from piling up.
THE FROST VEIL
Every tab the agent touches wears a visible marker. A WebGL frost sheet condenses in, breathes for the whole turn, and melts when the job is done. Move your mouse over it, and the light follows.
example.com
live WebGL demo · pointer moves the light · reduced-motion gets the static fallback
The indicator appears the moment the agent takes control and only leaves before the answer. It lives for the whole turn, never fading mid-action.
“Checking open tabs…”, “Clicking…”, “Typed into the search box”. Element text, never selectors. Machine errors get translated too.
Pointer-events pass through and the centre stays see-through, so you can keep reading while the agent works. No backdrop blur in the way.
WHAT IT’S FOR
Whatever you’d do to a page by hand (navigate, read, click, type), your agent can do, on the model you point it at.
“Log in and set these fields.” The agent types into the right boxes on the page you’re looking at, and you watch it happen.
“Summarise this page” or “grab the pricing table.” It snapshots real content, never inventing what isn’t there.
Search, open, compare, repeat. It keeps the conversation and your context across the whole flow, on one tab or several.
Point it at a page and ask it to check something. The veil stays up while it works, and melts when it reports back.
Run it entirely on a local model: private, offline, no API bill. Or hop to the cloud for a harder task.
Every chat is disposable by default. The one worth keeping? One tap files it into the Augmentor Chat workspace in DSH.
MEASURED, NOT MARKETED
"open example.com · snapshot · tell me the title" → prompt to idle in 3.42s on a local 27B model.
INSTALL
Two ways to install: do it by hand below, or copy one prompt and let your agent do it. If you already run DSH, you’re 90% of the way there.
Augmentor is a plugin for DeepSeek Harness (DSH), so DSH is the runtime you need first. Get it from github.com/deepseek-ai/deepseek-harness.
http://127.0.0.1:3080 and opens it in your browser. Once it’s up, come back and install Augmentor below.Run it from source instead: git clone https://github.com/deepseek-ai/deepseek-harness.git && cd deepseek-harness && pnpm install && pnpm run build && pnpm dsh web
chrome://extensions, turn on Developer mode.augmentor/extension/ folder.@deepseek-ai/dsh-augmentor into your profile.~/.dsh/cordis.patch.yml. The app live-watches that file and mounts the plugin without a restart.browser_* tools and its own GUI card (connection status, endpoint, dedicated directory).OR · AUTOMATIC INSTALL
Not in the mood to click around? Copy this prompt into the AI you already use — your DSH agent or any coding assistant. It installs DSH and Augmentor end-to-end and reports exactly what it did.
Works in an Augmentor chat too — it has the browser_* tools to open chrome://extensions and do the web steps for you.
FAQ
If you already run DeepSeek Harness with local or cloud models, you’re minutes away from browser control you can actually see.
Preview build · not yet on the store · load it unpacked from the repo.